Elasto-Net: An HDL Conversion Framework For Convolutional Neural Networks

Hardware solutions for Convolutional Neural Networks (CNN) have emerged in the recent wake of their success in image classification. Although CNNs are effective in classifying images, they can be highly complex to implement in hardware. CNNs come in many shapes and sizes. Thus, designing a universal hardware that is network agnostic is a tremendous challenge. In this paper, we propose to build a Hardware Descriptive Language (HDL) conversion framework that uses the Nested Processing Element(NPE) we designed [1] as its building block to spawn a Nested Processing Element based Computation Engine (NPECE). We implement an option called Elasto-Net that stretches the layer-agnostic property of the NPECE to the network level. Thus, this framework is able to realize a computation engine that is network agnostic for a specific hardware platform. The motivation to develop a network agnostic computation engine is to establish a convenient test bench infrastructure for testing the NPECE and make it available to higher level programmers. Thus Elasto-Net will help us test the effectiveness of NPECE on many combinations of architectures and CNNs such as AlexNet [2], GoogleNet [3] and ResidualNet [4]. Finally, we report the resource utilization for the layers that are constructed using Elasto-Net for the convolutional layers of AlexNet.